Ferroelectricity-modulated asymmetric van der Waals heterostructure for ultralow-power neuromorphic synapse and logic-in-memory operations

Why smarter, low-power chips matter

Everyday gadgets – from phones and cameras to smart speakers and home sensors – increasingly need to see, learn and react in real time. But today’s computer chips waste energy shuttling data back and forth between separate units for sensing, memory and processing. This paper presents a tiny, layered device that combines all three roles in one structure, dramatically cutting power use while still handling complex tasks like image recognition and brain-like learning.

Stacking ultra-thin materials into one tiny brain cell

The researchers build their device from several sheets of materials only a few atoms thick, stacked like a miniature club sandwich. The core is a special crystal called a ferroelectric, which can hold an internal electric “direction” that stays put even when power is removed. This layer sits atop other light-sensitive and conducting layers, with graphene at the bottom acting as a transparent, flexible contact. Because the layers only touch through weak van der Waals forces rather than traditional chemical bonds, they can be combined with great freedom, creating a highly tunable structure in a very small footprint.

Using built-in electric fields as a control knob

The key trick is to use the ferroelectric layer as an internal switch that reshapes how electrical charges move through the stack. By applying small positive or negative voltage pulses, the team can flip the direction of the ferroelectric’s internal field. That, in turn, raises or lowers the energy barriers at the interfaces between layers, changing how easily electrons can flow. Because this built-in field remains even after the pulse ends, the device naturally remembers its state without needing continuous power, much like a synapse in the brain remembers how strongly two neurons are connected.

Logic operations and artificial synapses in the same device

With this internal control, a single device can act as several different logic elements – the basic building blocks of digital circuits. By choosing the pulse pattern and how the output current is read, the authors implement five classic logic operations (AND, OR, NOT, NOR and NAND) all in one physical structure, rather than needing separate transistors and wiring for each gate. At the same time, by carefully engineering defects in one of the layers, the device behaves like a neuromorphic synapse: its conductance can be smoothly tuned across more than 128 distinct levels and adjusted by light or electrical pulses. These levels are stable, clearly separated from noise, and can be updated using vanishingly small energies, comparable to or even below those used in biological synapses.

Seeing and learning with light over a broad spectrum

Because some of the layers are light-sensitive, the device also functions as a high-performance photodetector. At zero applied voltage, it can sense light from ultraviolet to near-infrared while keeping its dark current – the background current with no light – at extremely low levels, which is crucial for detecting weak signals. When a small bias is applied, the same structure switches into a “photonic synapse” mode: bursts of light act like learning pulses, strengthening or weakening the effective connection in a way that mimics how real synapses respond over time. The team demonstrates behaviors such as short- and long-term memory, learning–forgetting–relearning cycles and classical conditioning, all driven directly by light.

From single device to intelligent vision systems

To show the practical impact, the authors build a conceptual image recognition system that uses many of these devices in parallel. In this design, the light-driven synaptic behavior captures and pre-processes visual features, while the reconfigurable logic behavior enhances and filters them in different ways. Combining these roles yields a recognition accuracy of about 97% on a standard image dataset, outperforming a system that relies on synaptic behavior alone. Overall, the work demonstrates a realistic path toward compact chips that can sense, remember and compute in place, opening the door to ultra-low-power cameras, smart sensors and neuromorphic vision hardware that operates much more like a biological eye–brain system than a conventional computer.

Citation: Zhi, J., Wen, Y., Chen, J. et al. Ferroelectricity-modulated asymmetric van der Waals heterostructure for ultralow-power neuromorphic synapse and logic-in-memory operations. Nat Commun 17, 3974 (2026). https://doi.org/10.1038/s41467-026-70668-w

Keywords: neuromorphic hardware, in-sensor computing, 2D materials, ferroelectric devices, image recognition